GZSL-Lite: A Lightweight Generalized Zero-Shot Learning Network for SSVEP-Based BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP) based brain-computer interfaces (BCIs) offer intuitive human-computer interaction.
• Generalized zero-shot learning (GZSL) networks aim to reduce user training burden in SSVEP BCIs by classifying unseen classes.
• Existing GZSL networks are computationally intensive, hindering practical BCI applications.

Purpose of the Study:

• To develop a lightweight GZSL network (GZSL-Lite) for SSVEP-based BCIs.
• To reduce the number of trainable parameters and training time in GZSL networks.
• To enhance the classification performance of SSVEP BCIs.

Main Methods:

• Proposed a dual-attention structure integrated into a GZSL framework (GZSL-Lite).
• Employed convolution-based networks for feature embedding of EEG signals, class templates, and sine templates, sharing network weights across stimulus frequencies.
• Utilized dual-attention mechanisms guided by class and sine templates to focus on relevant EEG features, followed by depthwise convolutional blocks for classification.

Main Results:

• GZSL-Lite achieved a significant reduction in trainable parameters, comprising less than 1% of state-of-the-art counterparts.
• The proposed network demonstrated substantial performance improvements in SSVEP classification accuracy.
• Experimental evaluations on two public datasets validated the efficacy and efficiency of GZSL-Lite.

Conclusions:

• The lightweight GZSL-Lite network effectively addresses the computational challenges of traditional GZSL models for SSVEP BCIs.
• Dual-attention mechanisms significantly enhance classification accuracy while minimizing model complexity.
• GZSL-Lite offers a practical and high-performance solution for user-friendly SSVEP-based brain-computer interfaces.